Michael Vácha1,2,3, Frank Hofheinz1, Anja Braune1,4,5, Steffen Löck2,3,6,7,8, Alex Zwanenburg3,7,8

1Institute of Radiopharmaceutical Cancer Research, Helmholtz-Zentrum Dresden-Rossendorf, Dresden, Germany

2Helmholtz-Zentrum Dresden-Rossendorf, Institute of Radiooncology - OncoRay, Dresden, Germany

3OncoRay-National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universität Dresden, Helmholtz-Zentrum Dresden-Rossendorf, Dresden, Germany

4Department of Nuclear Medicine, University Hospital Carl Gustav Carus, Technische Universität Dresden, Dresden, Germany

5Carl Gustav Carus Faculty of Medicine, Technische Universität Dresden, Dresden, Germany

6Department of Radiotherapy and Radiation Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universität Dresden, Dresden, Germany

7German Cancer Consortium, partner site Dresden, and German Cancer Research Center, Heidelberg, Germany

8National Center for Tumor Diseases (NCT), NCT/UCC Dresden, Dresden, Germany

Introduction

There is an increasing interest in using positron emission tomography (PET) data for segmentation, biomarker identification, and outcome prediction. To make PET values comparable between patients, these values are often normalized to a standard unit, which is, in most cases, the body-weight-normalized standardized uptake value (SUVbw). By definition, the SUVbw can be computed from the formula:

\[ \text{SUV}_{\mathrm{bw}} = \frac{\scriptstyle A_c \, W}{\scriptstyle D} \]

where \({A}_{c}\) represents the measured activity concentration within a region of interest or voxel in Bq/ml, \(W\) the weight of the patient in g, and \(D\) the total administered radionuclide dose at the time to which the voxel values correspond, in Bq. However, PET imaging data in Digital Imaging and Communications in Medicine (DICOM) files are usually not stored as standardized uptake values. Consequently, the voxel values and associated metadata have to be correctly interpreted by the viewing or analytical tools in order to convert the image units to SUVbw. In the radiomics field, many tools were not primarily designed for PET imaging and may not process all relevant metadata correctly. Consequently, identical images may be processed inconsistently across various software tools. This can give rise to two potential issues:

  • incorrect or inconsistent SUV calculation, resulting in under- or overestimation of SUVbw values, which may introduce unrecognized bias and reduce generalizability;
  • errors that lead to patient exclusion and a decreased sample size.

By creating reference standards for SUVbw computation by software tools, we aim to:

  • verify that SUVbw is computed correctly;
  • maximize the number of PET imaging series that can be used for quantitative image analysis.

Therefore, we:

  1. studied how PET images are stored in real-world imaging DICOM files;
  2. summarized the instructions on how to read these files and calculate SUVbw;
  3. generated a set of representative digital reference objects (DROs) to assess whether imaging software can accurately interpret diverse PET DICOM formats.




Methods

Metadata analysis

To analyze how PET imaging data are stored, we searched for real-world PET human imaging series via two sources:

We identified 37 cohorts containing relevant PET imaging data: ACRIN-FLT-Breast, ACRIN-NSCLC-FDG-PET, Anti-PD-1_Lung, BREAST-DIAGNOSIS, CC-Tumor-Heterogeneity, CMB-CRC, CMB-GEC, CMB-LCA, CMB-MEL, CMB-MML, CMB-PCA, CPTAC-CM, CPTAC-HNSCC, CPTAC-LSCC, CPTAC-LUAD, CPTAC-PDA, CPTAC-SAR, CPTAC-UCEC, CT-vs-PET-Ventilation-Imaging, Head-Neck-PET-CT, Internal-Berlin, Internal-Dresden, Internal-Munich, Lung-PET-CT-Dx, NaF PROSTATE, NSCLC Radiogenomics, QIN-BREAST, RIDER Lung PET-CT, Soft-tissue-Sarcoma, TCGA-BLCA, TCGA-KIRP, TCGA-LUAD, TCGA-LUSC, TCGA-PRAD, TCGA-THCA, TCGA-UCEC, and VAREPOP-APOLLO.

We excluded imaging series without attenuation correction and detector normalization applied, as well as maximum intensity projections and other reprojections. For all remaining series, we scanned a preselected set of attributes (listed in Suppl. Table 1) and analyzed how they were used in practice.


Manual for PET DICOM SUVbw conversion

Based on literature (DICOM standards, QIBA consensus, Turku PET center manual), our expertise, and the metadata analysis, we summarized the rules for converting PET images into SUVbw-normalized images.


Construction of digital reference objects

Furthermore, we assembled a comprehensive set of digital reference objects (DROs) for verifying SUV conversion. All DRO DICOM files were synthesized de novo using Python pydicom library version 3.0.1. The design of these DROs was chosen to resemble common PET calibration phantoms – each DRO includes one hot sphere (SUVbw = 4.00), one cold sphere (SUVbw = 0.20), a background region (SUVbw = 1.00), and surrounding zero-activity region (SUVbw = 0.00) (See Fig. 1).

All DROs are identical in terms of the design, volumes, and resulting SUVbw values, while the stored voxel values and DICOM attributes vary based on the intended use of each DRO. The binary mask for feature extraction covers the whole DRO volume, excluding the surrounding region. To verify a software tool computes the SUVbw correctly, either:

  1. convert the DRO DICOM file to SUVbw and extract the DRO maximum, minimum, and median SUVbw values from the region defined by the mask, or

  2. convert the DRO DICOM file to SUVbw and visually inspect the values in the hot sphere, cold sphere, and background region.

All values should be calculated with a precision of two decimal digits. The target values are identical across all cases and are defined as follows:

  • Hot sphere / maximum DRO value = 4.00 SUV
  • Cold sphere / minimum DRO value = 0.20 SUV
  • Background region / median DRO value = 1.00 SUV


Fig. 1: Visualization of the DRO design in axial, coronal, and sagittal planes (from left to right). Voxel values are shown in SUVbw.

Fig. 1: Visualization of the DRO design in axial, coronal, and sagittal planes (from left to right). Voxel values are shown in SUVbw.




Results

Baseline characteristics

For the metadata analysis, we identified and examined 3433 PET imaging series from 1775 patients. These images were acquired by a large variety of PET scanner models constructed by 3 major PET scanner manufacturers (See Table 1).


Table 1: Percentage of series and a list of unique ManufacturerModelName (0008,1090) values by scanner manufacturer. Siemens, CTI, and CPS were merged for the sake of simplicity.
Manufacturer Perc Models
GE 59% Advance, Discovery 610, Discovery 690, Discovery 710, Discovery IQ, Discovery LS, Discovery MI, Discovery MI DR, Discovery RX, Discovery ST, Discovery STE
Philips 11% Allegro Body(C), GEMINI TF Big Bore, GEMINI TF TOF 16, Guardian Body(C), TruFlight Select
Siemens/CTI/CPS 29% 1023, 1024, 1062, 1080, 1093, 1094, 962, Biograph 20_mCT, Biograph 64_mCT, Biograph Horizon, Biograph_mMR, Biograph128_mCT 4R, Biograph128_Vision 450 Edge, Biograph16_Horizon 3R, Biograph20_mCT, Biograph20_mCT 3R, Biograph40_mCT, Biograph40_mCT 4R, Biograph40_TruePoint, Biograph6_TruePoint, Biograph64_mCT, Biograph64_mCT 3R, Biograph64_Vision 600, Somaris/5 3D, Somaris/5 3D Postprocessing, SOMATOM Definition AS_mCT
Unknown 1% DicomCleaner, Integrity Medical Image Importer



Rescale slope (m) and intercept (b)

Background

Because of the wide range of the measured voxel values in PET imaging data, the use of rescale slope (RescaleSlope; 0028,1053) ensures that stored values are within the range that can be stored in the voxel data type (e.g., signed 16-bit integers can store integer values from -32,768 to 32,767). Applying the rescale slope to the stored voxel data is the mandatory first step in SUV calculation, independent of the units encoded in the DICOM image:

\[ \text{SUV}_{bw} = \frac{\scriptstyle(m \, P + b)\, W}{\scriptstyle D} \]

where P is the stored voxel value (PixelData; 7fe0,0010) in Bq/ml, m the rescale slope, b the rescale intercept (RescaleIntercept; 0028,1052). The rescale slope may vary for each slice and must be applied on a slice-wise basis. Since the rescale intercept is required to be 0 in all PET studies, it can be omitted from the formula:

\[ \text{SUV}_{bw} = \frac{\scriptstyle m \, P \, W}{\scriptstyle D} \]


Metadata analysis

Comparing series, a wide range of rescale slopes was observed in the studied data.

In general, the rescale slope values within each series were either (see Table 2):

  • identical
  • different in each slice
  • rather constant with a few exceptions (usually representing slices with the highest uptake, e.g., the tumor region).


Table 2: Comparison of unique RescaleSlope (0028,1053) values and number of slices per series.
N unique RescaleSlope values Freq Perc
= 1 1342 39 %
1 < N_Slices 391 11 %
= N_Slices 1698 49 %


In our dataset, all PET imaging series had RescaleIntercept = 0.


Digital reference object

The following DRO was constructed:

  • DRO with multiple values of the RescaleSlope attribute.

    Possible issues:

    • RescaleSlope is not applied slice-wise


Recommendations

RescaleIntercept (0028,1052) attribute should be checked to ensure it equals 0. Otherwise, exclude the dataset.

The corresponding RescaleSlope (0028,1053) has to be applied to all stored voxel values within a slice, independently of other parameters such as the units.



Voxel values (P) and their unit

Background

The attribute PixelData (7fe0,0010) stores the voxel values. As for other imaging modalities, the PixelData attribute can be read using other attributes from the DICOM Image Pixel Module, such as Rows (0028,0010), Columns (0028,0011), Bits Allocated (0028,0100), or Pixel Representation (0028,0103). While single-slice DICOM images still prevail, multi-planar DICOM images containing 3-D voxel data (i.e., ‘one-file DICOM’ formats) may become increasingly common in the future. For PET imaging, the interpretation of these values depends on the unit specified in the Units (0054,1001) attribute, which is explained in the following subsections.


Units = “BQML”

Voxel values are commonly in becquerels per milliliter: the attribute Units (0054,1001) is set to “BQML”. Assuming W and D are accordingly corrected and expressed in proper units, P may be used directly in the formula:

\[ \text{SUV}_{\mathrm{bw}} = \frac{\scriptstyle m \, P \, W}{\scriptstyle D} \]

However, other units may be used by scanners or due to previous processing of the DICOM file.


Units = “GML”

The voxel values can be in grams per milliliter (Units = GML). This indicates the activity concentration was already normalized. In most cases, the normalization corresponds to body-weight normalization (BW), where the SUV voxel value equals the stored value multiplied by the rescale slope.

\[ \text{SUV}_{bw} = m \, P \]

However, there may be scenarios where other normalization techniques were applied. Specifically, four SUV methods are compliant with the current DICOM standard and result in the unit GML. The method used can be extracted from the attribute SUVType (0054,1006) and should be considered as BW if empty. The following factors are used, as reported by Sugawara:

  • lean body mass by Morgan (SUVType = “LBM”)

    for male patients: \(\text{LBM} = 1.10 W - 120 ({\scriptstyle \frac{W}{H}})^2\)

    for female patients: \(\text{LBM} = 1.07 W - 148 ({\scriptstyle \frac{W}{H}})^2\)

  • lean body mass by James et al. (James, William Philip Trehearne, and J. C. Waterlow. Research on obesity. 1976) / Morgan (SUVType = “LBMJAMES128”)

    for male patients: \(\text{LBM} = 1.10 W - 128 ({\scriptstyle \frac{W}{H}})^2\)

    for female patients: \(\text{LBM} = 1.07 W - 148 ({\scriptstyle \frac{W}{H}})^2\)

  • lean body mass by Janmahasatian (SUVType = “LBMJANMA”)

    \(\text{BMI} = \scriptstyle \frac{W}{H^2}\)

    for male patients: \(\text{LBM} = \scriptstyle \frac{ 9270 W}{6680 + 216 \text{BMI}}\)

    for female patients: \(\text{LBM} = \scriptstyle \frac{9270 W}{8780 + 244 \text{BMI}}\)

  • ideal body weight (SUVType = “IBW”)

    for male patients: \(\text{IBW} = 48.0 + 1.06 (H - 152)\)

    for female patients: \(\text{IBW} = 45.5 + 0.91 (H - 152)\)

For all formulas, W is weight in kg, and H is height in cm. Additionally, the following attributes are required:

  • PatientSize (0010,1020) - Height of the patient in m

  • PatientSex (0010,0040) - Sex of the patient - M (male), F (female) or O (other)

In turn, these factors can be used for the normalization, e.g., for SUVlbm:

\[ \text{SUV}_{\mathrm{lbm}} = \frac{\scriptstyle A_c \, \text{LBM} \, 10^{3}}{\scriptstyle D} \]

To get the SUVbw values, the stored voxel values can be divided by the corresponding factor and multiplied by the patient’s weight and the rescale slope of the slice, e.g., for LBM:

\[ \text{SUV}_{bw} = \frac{\scriptstyle m \, P \, W}{\scriptstyle \text{LBM} \, 10^{3}} \]

There is currently no consensus on the computation of LBM and IBW when PatientSex is specified as “O”. Using the mean value of the sex-specific factors appears to be a reasonable and simple solution.


Units = “CM2ML”

The unit square centimeters per milliliter (Units=“CM2ML”) corresponds to images normalized to body surface area (BSA) using the Du Bois formula:

\[ \text{BSA} = 0.007184 \, H^{0.725} \, W^{0.425} \]

where H is height in cm, and W is weight in kg. The SUVType (0054,1006) should be set to “BSA” in this case. SUVbw can be backcomputed analogically to the previous section:

\[ \text{SUV}_{bw} = \frac{\scriptstyle m\, P \, W}{\scriptstyle \text{BSA} \, 10^{4}} \]


Units = “CNTS”

Philips scanners often save voxel values as counts (Units=“CNTS”). In this case, scale factors may be provided for converting the stored values to activity concentration or directly to SUV values. SUV scale factor allows direct conversion of the voxel values to SUVbw. The formula is mentioned in Philips conformance statements (here, for Philips Ingenuity):

(7053,1000) DS SUV Scale Factor: This value only applies when Units (0054,1001) is equal to CNTS. The SUV Scale Factor is used to convert the voxel data from counts to an SUV value. This is done by using the formula SUV Value = ((SV x m) + b) x f, where: SV = original stored voxel value, m = Rescale Slope (0028,1053), b = Rescale Intercept (0028,1052), f = SUV Scale Factor (7053, 1000). If the SUV Scale Factor is 0.0, then the voxel data cannot be converted from counts to an SUV value.

The Activity Concentration Scale Factor allows the conversion of the voxel values to Bq/ml, as explained in the conformance statement (here, for Philips Ingenuity):

(7053,1009) DS Activity Concentration Scale Factor: This value only applies when Units (0054,1001) is equal to CNTS. The Activity Concentration Scale Factor is used to convert the voxel data from counts to Activity Concentration (in Bq/ml). This is done by using the formula Activity Concentration Value = ((SV x m) + b) x f, where: SV = original stored voxel value, m = Rescale Slope (0028,1053), b = Rescale Intercept, (0028,1052), f = Activity Concentration Scale Factor (7053, 1009). If the Activity Concentration Scale Factor is 0.0, then the voxel data cannot be converted from counts to Activity Concentration.

In case none of the factors is provided, counts (CNTS) can be converted to counts per second (CPS) by dividing by the actual frame duration in seconds, i.e., the Actual Frame Duration attribute (0018,1242) value divided by 1000.


Units = “CPS”

Counts per second (CPS) can be converted to activity concentration (Bq/ml) based on whether the image is dose calibrated or not, which can be derived from the presence of “DCAL” in the attribute Corrected Image (0028,0051):

  • in case the image is already dose calibrated, by dividing by voxel volume in ml;
  • in case the image is not dose calibrated, by multiplying by the dose calibration factor and dividing by voxel volume in ml.

The dose calibration factor is stored in the corresponding DICOM attribute Dose Calibration Factor (0054,1322). However, in most cases, the factor may be unknown. The voxel volume can be computed from the Pixel Spacing (0028,0030) and Slice Thickness (0018,0050) attributes, that show the dimensions of a single voxel in mm.

Furthermore, it must be verified that all required image corrections have been applied, as this unit is often associated with uncorrected image data.


Other units

For other units, the conversion to activity concentration or SUVbw is impossible or unclear.

That includes, above all, following units:

  • PROPCNTS - proportional to counts, not mentioned in conformance statements, conversion is not clear

  • PROPCPS - proportional to counts per second, mentioned by Siemens for “non-quantitative, non-attenuation corrected images”, conversion is not clear

  • 1CM - 1/centimeter, used for MU maps


Metadata analysis

The units used were highly dependent on scanner manufacturer (see Table 3):

Table 3: Frequency of unique Units (0054,1001) value by manufacturer
BQML CNTS CPS GML PROPCNTS
Total 2987 248 121 44 33
GE 1839 0 121 34 33
Siemens/CTI/CPS 1000 0 0 0 0
Philips 113 248 0 10 0
Unknown 35 0 0 0 0

There were no cases with a non-NA value in the SUVType attribute.

Regarding counts as units and their conversion, 90% of the series with counts units had at least one scale factor provided, of which the SUV scale factor appeared most often. The activity concentration scale factor never appeared without the SUV scale factor (See Table 4).

Table 4: Proportions of images with ‘CNTS’ units, showing whether Activity Concentration Scale Factor (ACSF; 7053,1009) and/or SUV Scale Factor (SUVSF; 7053,1000) are present in the metadata.
ACSF SUVSF Freq Perc
TRUE TRUE 136 55 %
FALSE TRUE 87 35 %
FALSE FALSE 25 10 %
TRUE FALSE 0 0 %


Digital reference object

We synthesized multiple DROs verifying a standardized conversion to SUVbw:

  • DRO with Units = BQML and Decay Correction = START (baseline DRO)

    Possible issues:

    • unit BQML is not implemented.
  • DRO Units = GML (corresponding to SUVbw)

    Possible issues:

    • unit GML is not implemented;

    • SUVtype is required (while should be considered “BW” by default);

    • RescaleSlope is not applied;

    • other corrections are applied.

  • DRO Units = GML (corresponding to SUV LBMJAMES128)

    Possible issues (not considering previously mentioned):

    • SUVlbm is not implemented;

    • incorrect formula for LBM is used (instead of LBMJAMES128 DICOM standard);

    • LBM unit is not changed to g;

    • patient’s weight, height or sex are not extracted correctly;

    • SUV type is not extracted correctly.

  • DRO PatientSex = “O” (other) with Units = GML (corresponding to SUV IBW)

    Possible issues (not considering previously mentioned):

    • SUVibw is not implemented;

    • strategy for PatientSex = “O” is not implemented.

  • DRO Units = CM2ML (corresponding to SUVbsa)

    Possible issues (not considering previously mentioned):

    • SUVbsa is not implemented;

    • incorrect formula for BSA is used;

    • BSA unit is not changed to cm2.

  • DRO Units = CNTS using Philips SUV scale factor

    Possible issues (not considering previously mentioned):

    • unit CNTS is not implemented;

    • SUV scale factor is not extracted or applied correctly.

  • DRO Units = CNTS using Philips activity scale factor

    Possible issues (not considering previously mentioned):

    • activity scale factor is not extracted or applied correctly.


Recommendations

tbd



Patient’s weight (W)

Background

PatientWeight (0010,1030) is often manually entered, which can result in missing values or incorrect units. The correct DICOM unit is kilograms, which must be converted to grams for SUV computation. Values exceeding 1000 indicate that the weight was entered in grams. Furthermore, patient weight is occasionally incorrectly stored in another DICOM attribute, particularly PatientSize (0010,1020).


Metadata analysis

In our dataset, no PatientWeight values exceeded 1000. In 0.8% of cases, the value was missing.


Digital reference object


Recommendations

PatientWeight (0010,1030) attribute should be checked that it is present and != 0. Otherwise, SUV can’t be computed.

PatientWeight (0010,1030) attribute should be checked to ensure that the value was entered in kilograms. If the value > 1000, divide by 1000.



Radiopharmaceutical dose (D)

Background

The RadionuclideTotalDose (0018,1074) DICOM field records the total administered dose of the radiopharmaceutical. It is another attribute that is typically recorded manually, which may result in missing values or incorrect units.
Furthermore, the unit may not be standardized as different Information Object Definitions (IODs) allow either Bq or MBq. For the equation, the unit has to be recognized by its value and converted to Bq.

For SUV conversion, the administered dose must correspond to the time at which the voxel values occurred. This requires a correction of the administered dose for radioactive decay of the radionuclide between the time of administration and the corresponding reference time point. Therefore, it is essential to know whether the images were decay-corrected and, if so, to which time point. This information is provided by the DecayCorrection (0054,1102) attribute, which can take one of three values (“ADMIN,” “START,” or “NONE”). As shown below, the subsequent correction steps depend on this attribute.

Decay correction = ADMIN

The value “ADMIN” indicates that the PET image data were decay-corrected to the time of radipharmaceutical administration. In this case, no additional correction of the administered dose is required, because it already represents the dose at the time point to which the images were decay-corrected. The formula can be written as:

\[ \mathrm{SUV}_{bw} = \frac{\scriptstyle m \, P \, W}{\scriptstyle D_\mathrm{adm}} \]

Decay correction = START

In most cases, the decay correction attribute is set to “START”, implying the series was decay corrected to a reference time representing the start of the PET image acquisition. The administered dose has to be adjusted to account for decay between administration and the reference time, using the formula: \(D = {D}_{adm} e^{-\lambda (t_\mathrm{ref}-t_\mathrm{adm})}\). Then, SUV can be computed using the formula:

\[ \mathrm{SUV}_{bw} = \frac{\scriptstyle m \, P \, W}{\scriptstyle D_\mathrm{adm} \, e^{\left(-\lambda (t_\mathrm{ref}-t_\mathrm{adm}) \right)}} \]

where \(D_{adm}\) the administered dose of the radionuclide (0018,1074) in Bq, \(t_{ref}\) the reference time, \(t_{adm}\) the radiopharmaceutical administration time (0018,1072). \(\lambda\) represents the decay constant for the radionuclide and is computed as \(\lambda = \frac{\ln(2)}{T_{1/2}}\) where \(T_{1/2}\) is the radionuclide half life (0018,1075).

Determining the reference time:

The reference time, to which the image was corrected can be derived by multiple methods.

Option A (QIBA):

The Quantitative Imaging Biomarkers Alliance (QIBA) recommends the following approach (ordered by priority):

  1. Use Series Date (0008, 0021) and Series Time (0008, 0031) if not modified during post-processing (in that case, Series Time would be later than Acquisition Time).
  2. Use GE private tag (0009,100D) if present (PET Scan DateTime).
  3. Alternatively, use the earliest Acquisition Date and Time (0008,0022) and (0008,0032) - which is mentioned as potentially “wrong for case of PETsyngo 3.x multi-injection”.
  4. Ultimately, backcompute from acquisition time, frame reference time, and frame duration, using the formula:

\[ \text{t}_{ref} = \text{t}_{acq} + \text{T}_{ave} - {\Delta{t}} \]

where \({t}_{acq}\) is the acquisition time, \(\Delta{t}\) is the frame reference time in seconds, and \({T}_{ave}\) is the average count rate time in seconds which is implementation-dependent. One of the common formulas is: \({T}_{ave} = \frac{1}{\lambda} \text{ln} \frac{(\lambda \text{T})}{1 - e^{-\lambda \text{T}}}\) where \(\lambda = \frac{\ln(2)}{T_{1/2}}\), \(\Delta{t}\) is the frame reference time in sec, and \(T\) is the frame duration in sec, which leads to the formula:

\[ \text{t}_{ref} = \text{t}_{acq} + \frac{1}{\scriptstyle \lambda} \text{ln} \frac{\scriptstyle (\lambda \text{T})}{\scriptstyle 1 - e^{-\lambda \text{T}}} - {\Delta{t}} \]

Ad 2) GE states in their conformance statements:

  • PET scan_datetime (0009,100D; DT)
  • Series Date (0008,0021; DA): Extract date from Scan.scan_datetime
  • Series Time (0008,0031; TM): Extract time from Scan.scan_datetime

This is only stated for some models (e.g., Discovery ST/ RX/ STE) unlike some other models (e.g., Discovery 710/610 or Optima 560). Furthermore, in the 2009 response to QIBA by GE, this private attribute is mentioned to be used for dose correction in case of processed images, where SeriesTime > AcquisitionTime.

Siemens also uses a private tag as can be seen in its conformance statements:

  • Decay Correction DateTime (0071,1022; DT): The date and time to which the image was decay corrected. Also refer to (0054,1102)
  • Series Date (0008,0021; DA): The time in which acquisition was initiated. Note: this is also a time used for decay correction (see attribute: (0054,1102))
  • Series Time (0008,0031; TM): See above
  • Decay Correction (0054,1102; CS): The real-world event to which images in this Series were decay corrected. Value: START. This refers to the Series Date and Time, see (0008,0021) and (0008,0031) and private attribute (0071,1022)

This suggests that the GE private tag overwrites SeriesDate and SeriesTime and should therefore be preferred when present, whereas the Siemens private tag appears to be irrelevant for our purposes.

Option B (our proposal):

We suggest always using the method 4) – backcomputation from acquisition time, frame reference time, and frame duration. This method seems to be the most robust for research use since:

  • the method is independent of potential modifications of SeriesTime during anonymization or post-processing;
  • the method does not rely on private tags that may not be present or removed during anonymization;
  • the method does not rely on all slices being presented in the series (earliest AcquisitionTime may not represent the start of the acquisition for cropped images);
  • the method can be applied to all image series regardless of scanner type or manufacturer.

If one of the required attributes is missing, the series should be excluded from the analysis.


Decay correction = NONE

Another scenario involves PET images that have not been decay-corrected, with DecayCorrection set to “NONE”. In this case, more complex formulas must be applied to ensure that each image is corrected to a reference timepoint, that matches the one used for the administered dose. This is particularly challenging in multi-frame or dynamic series consisting of images with different acquisition times. To perform this correction, the voxel values must be scaled to the reference time using the factor \(e^{\lambda (t_{\text{scan}} - t_{\text{ref}})}\) where the \({t}_{scan}\) is derived from the acquisition time and the average count rate time \({t}_{scan} = {t}_{acq} + {T}_{ave}\). Using the \({T}_{ave}\) expression from the previous section, this term can be incorporated into the original equation, resulting in the following formula:

\[ \mathrm{SUV}_{bw} = \frac{ m \, P \, W }{ D_{\text{adm}} } \frac{ e^{\lambda ({t}_{acq} + \frac{1}{\lambda} \text{ln} \frac{(\lambda \, \text{T})}{1 - e^{-\lambda \text{T}}} - t_{\text{ref}})} }{ e^{\! -\lambda (t_{\text{ref}} - t_{\text{adm}})} } \]

That can be simplified to:

\[ \mathrm{SUV}_{bw} = \frac{\scriptstyle m \, P \, W}{\scriptstyle D_{\mathrm{adm}}} \frac{\scriptstyle \lambda \, T}{\scriptstyle 1 - \mathrm{e}^{-\lambda T}} \mathrm{e}^{\lambda (t_{\mathrm{acq}} - t_{\mathrm{adm}})} \]

where the SUVbw can be computed for each image directly from the values in the DICOM attributes. This method may lead to minor inaccuracies due to different implementations of the average count rate time \({T}_{ave}\) between scanner manufacturers.


Metadata analysis

After excluding non-fluorine-18 series, two distinct peaks at approximately 4 × 10² and 4 × 10⁸ suggest that the administered dose was stored in MBq or Bq, respectively (see Fig. 2):

Fig. 2: Histogram of RadionuclideTotalDose (DICOM tag 0018,1074) values for FDG-PET series.

Fig. 2: Histogram of RadionuclideTotalDose (DICOM tag 0018,1074) values for FDG-PET series.

Regarding the dose correction, most series were decay-corrected to the acquisition start time (DecayCorrection = “START”), while there were no series decay-corrected to the administration time (DecayCorrection = “ADMIN”, Table 5):

Table 5: Frequency of DecayCorrection (0054,1102) values by manufacturer.
DC GE Philips Siemens/CTI/CPS Unknown
START 2022 279 1000 35
NONE 5 92 0 0

More than a half of the series contained multiple AcquisitionTime (0008,0032) values (Table 6):

Table 6: Comparison of the number of unique AcquisitionTime (0008,0032) values per series and the number of slices.
N unique AcquisitionTime values Freq
= 1 1393
1 < N_Slices 2040
= N_Slices 0


The GE scan datetime often differed from the SeriesTime (Table 7):

Table 7: Comparison of GE private PET scan_datetime (0009,100D) and SeriesTime (0008,0031).
SeriesTime AcquisitionTime
= private tag 385 289
!= private tag 128 224
no_value 0 0


While, the Siemens scan datetime was always identical to SeriesTime (Table 8):

Table 8: Comparison of Siemens private Decay Correction DateTime (0071,1022) and SeriesTime (0008,0031).
SeriesTime AcquisitionTime
= private tag 166 5
!= private tag 0 161
no_value 0 0


Digital reference object

Following DROs for decay correction alternatives were created:

  • DRO dose in MBq

    Possible issues (not considering previously mentioned):

    • the dose is considered to be in Bq.
  • DRO DC = ADMIN

    Possible issues (not considering previously mentioned):

    • DecayCorrection == ADMIN is not implemented.
  • DRO DC = START but SeriesTime after AcquisitionTime (DC = START, option 4)

    Possible issues (not considering previously mentioned):

    • SeriesTime is not compared with AcquisitionTime;

    • the reference time formula is not implemented;

    • administered dose is not corrected to the same reference time among frames.

  • DRO GE private DC datetime (DC = START, option 2)

    Possible issues (not considering previously mentioned):

    • GE PET scan_datetime is not implemented.
  • DRO DC = NONE + multiple values ACQ TIME

    Possible issues (not considering previously mentioned):

    • DecayCorrection == NONE is not implemented;

    • AcquisitionTime and FrameDuration are not extracted correctly;

    • wrong formula for voxel value decay correction and dose correction is used;

    • the voxel values and the administered dose are not corrected to the same timepoint.


Recommendations

tbd



Radiopharmaceutical start time (tadm)

Background

Originally, radiotracer administration time was stored in the DICOM tag RadiopharmaceuticalStartTime (0018,1072). However, this attribute lacks the information about the date, therefore, was deprecated and Radiopharmaceutical Start DateTime (0018,1078) should be used instead.


Metadata analysis

Among the examined data, 59 % of series included only the RadiopharmaceuticalStartTime without an associated date, 35 % included both values, and 6 % included neither (Table 9). There was only 1 case in which only RadiopharmaceuticalStartDateTime was present, which may be due to the fact that the other attribute was deprecated only recently.

Table 9: Presence of Radiopharmaceutical Start Time (0018,1072, ‘RPSTime’) and Radiopharmaceutical Start DateTime (0018,1078, ‘RPSDateTime’) attributes.
RPSTime RPSDateTime Freq Perc
TRUE FALSE 2030 59 %
TRUE TRUE 1192 35 %
FALSE FALSE 210 6 %
FALSE TRUE 1 0 %


Digital reference object

Following objects were created:

  • DRO only Radiopharmaceutical Start Datetime, no Radiopharmaceutical Start time

    Possible issues:

    • Radiopharmaceutical Start Time is required.
  • DRO only Radiopharmaceutical Start Time, no Radiopharmaceutical Start Datetime

    Possible issues:

    • Radiopharmaceutical Start Datetime is required.
  • DRO over midnight (started before midnight, ended after midnight, only Radiopharmaceutical Start Time provided)

    Possible issues:

    • administration time and scan time are not compared.


Recommendations

tbd



Radiopharmaceutical half-life (T1/2)

Background

In the SUV conversion, it has to be accounted for different tracers – i.e., different radionuclide half-lifes, stored in the attribute RadionuclideHalfLife (0018,1075) in seconds. Alternatively, e.g., in case of missing value, the half-life can be determined based on RadionuclideCodeSequence (0054,0300) that stores the radionuclide codes.


Metadata analysis

All series had the Radionuclide half-life provided. Most series used Fluorine‑18 (F-18) as radionuclide (Table 10).

Table 10: Unique values of RadionuclideHalfLife (0018,1075); half-life values were truncated.
RadionuclideHalfLife (s) Freq Perc Corresponds to
598 13 0 % N-13
1223 2 0 % C-11
4057 20 1 % Ga-68
6586 1247 36 % F-18
6588 2081 61 % F-18
23400000 65 2 % Ge-68


Digital reference object

To check that these attributes are taken into account when SUV is computed, following DROs were constructed:

  • DRO with Ga-68 as radionuclide, half-life provided

    Possible issues:

    • half-life is not extracted correctly from RadionuclideHalfLife;
    • radionuclide half-life is not considered within the dose decay correction.


Recommendations

tbd




Discussion (tbd)

In the past, efforts have been made to standardize SUV computation, most notably by the Quantitative Imaging Biomarkers Alliance (QIBA). Furthermore, digital reference objects for SUV computation have been developed; however, these only evaluate the most common scenario with Units = BQML and DecayCorrection = START.

The present manual, together with the extended set of DROs, broadens the scope to cover multiple acquisition and metadata scenarios and can therefore support consistent and reproducible implementation of SUV conversion across software platforms and institutions – an aspect that is particularly critical in multi-center studies.

A limitation of this work that it solely focuses on the body-weight normalized SUV (SUVbw). Other commonly used SUV normalizations, such as SUVlbm, SUVbsa, and SUVibw are not explicitly addressed; however, these can be computed analogously by replacing body weight with the corresponding normalization factors described above.



Supplementary material


Suppl. Table 1: DICOM attributes scanned in the metadata analysis.
DICOMTag Attribute Vendor-specific
0008,0021 SeriesDate No
0008,0022 AcquisitionDate No
0008,002A AcquisitionDateTime No
0008,0031 SeriesTime No
0008,0032 AcquisitionTime No
0008,0060 Modality No
0008,0070 Manufacturer No
0008,103E SeriesDescription No
0008,1090 ManufacturerModelName No
0009,100D GEDecayCorrectionDateTime GE
0009,103B GEAdministrationDateTime GE
0010,0010 PatientName No
0010,0020 PatientID No
0010,0040 PatientSex No
0010,1020 PatientSize No
0010,1030 PatientWeight No
0018,1072 RadiopharmaceuticalStartTime No
0018,1074 RadionuclideTotalDose No
0018,1075 RadiotracerHalfLifeTime No
0018,1078 RadiopharmaceuticalStartDateTime No
0018,1242 ActualFrameDuration No
0018,9701 DecayCorrectionDateTime No
0028,1052 RescaleIntercept No
0028,1053 RescaleSlope No
0054,0300 RadionuclideCodeSequence No
0054,1000 SeriesType No
0054,1001 Units No
0054,1006 SUVType No
0054,1102 DecayCorrection No
0054,1300 FrameReferenceTime No
0054,1321 DecayFactor No
0054,1322 DoseCalibrationFactor No
0071,1022 SiemensDecayCorrectionDateTime Siemens
7053,1000 PhilipsSUVScaleFactor Philips
7053,1009 PhilipsActivityConcentrationScaleFactor Philips